Perplexity Revolutionizes AI Search with 'Search as Code'

Perplexity has unveiled a major innovation in the field of artificial intelligence research with its 'Search as Code' feature. This new architecture allows AI models to generate their own research workflows in the form of Python code, thus replacing fixed API search calls. This approach promises more accurate results while reducing token usage.

Traditionally, AI agents follow a repetitive pattern: writing a query, receiving a list of results from an API, reading those results, and then reformulating the query. This often inefficient process is referred to as a bottleneck by Perplexity. Current search engines, designed for human users, are not suited to the needs of AI agents that perform hundreds of searches in a short amount of time.

An Innovative Architecture

Perplexity's solution is based on a three-layer architecture:

• Model: it understands the task and develops the search strategy.
• Sandbox: a secure environment where the code is executed.
• Agentic Search SDK: it breaks down the search engine into individual functions that the model can combine.

Traditional search APIs remain available for simple queries, but for complex tasks, 'Search as Code' enables parallel searches and programmatic filtering of results, thereby optimizing the context window.

A Concrete Application in Cybersecurity

To demonstrate the effectiveness of this technology, Perplexity tested 'Search as Code' on a complex cybersecurity task. An agent was required to identify 200 critical software vulnerabilities (CVE) published between 2023 and 2025, finding for each the official vendor advisory, the affected software, and the corrective version. News articles or blogs were not counted. The model devised a three-step script, tailoring its searches to the specific formats of security bulletins from vendors like Mozilla and Google, and successfully completed the task with 85% fewer tokens than a standard pipeline. Competing systems achieved less than a quarter of the correct data.

Performance and Prospects

Across all of Perplexity's benchmarks, 'Search as Code' ranks first in four out of five categories and is nearly tied with OpenAI only on HLE. A recent study revealed that popular search agents often cheat on benchmarks like BrowseComp. When tested on a new benchmark with recent facts, each system saw its score drop by 25 to 40 points. 'Search as Code', currently deployed in Perplexity Computer and the Agent API, could address these limitations by offering a more accurate and efficient approach to AI research.